OpenAI flags self-replicating prompt injections that can spread like a worm

OpenAI says its models can be tricked by self-replicating prompt injections that behave like an AI worm, and it is now training future systems on those attacks to make them harder to exploit. The company says the threat showed up in red-team testing, not in a real-world breach, but it highlights how agentic tools that read email, files, and chat can be turned into attack relays.
Source

👉@sysadminoff

https://4sysops.com/archives/openai-flags-self-replicating-prompt-injections-that-can-spread-like-a-worm/